Himanshu H-lamba

`whoami`

name:        Himanshu
role:        B.Tech CSE (AI & ML) @ Vellore Institute of Technology  ·  CGPA 9.01/10
focus:       Distributed systems · Backend internals · Applied ML · Causal measurement
currently:   Writing C++17 concurrency primitives and LLM evaluation harnesses
philosophy:  "A benchmark you didn't run is a claim you can't defend."
reach_me:    hlamba24191@gmail.com

I build systems from the primitive up — a DNS resolver that hand-encodes the binary wire format over raw UDP sockets rather than importing a library, an Operational Transform engine proven across 30,000 randomized multi-client scenarios, a cache whose lock contention I found with a profiler before I fixed it with striping.

The thread running through all of it: measure the thing, then say the thing. Every number on this profile came out of a benchmark, a test suite, or a confidence interval — not a README written after the fact.

Arsenal

Languages

Systems & Backend

Machine Learning & Generative AI

Practices & Cloud

The unbadgeable half — click to expand

Domain	What I actually work with
Distributed Systems	Consistent hashing · Sharding · Replication · Heartbeat failure detection · Eventual consistency · Operational Transform · Cache-line-aligned lock striping
Concurrency	POSIX threads · Mutex contention profiling · ThreadSanitizer · Lock-free reasoning · False-sharing elimination
Networking	Raw UDP/TCP sockets · Binary wire-format encoding · WebSockets · Iterative & recursive resolution · REST API design
Data Engineering	Composite B-tree indexing · Query optimization · Schema design · Append-only logs · Compare-and-set versioning · TTL-aware caching
Applied ML	QLoRA / PEFT fine-tuning · RAG pipelines · FAISS vector retrieval · FinBERT sentiment · LSTM forecasting · Gradient boosting · Probability calibration
Measurement Science	Randomized holdout experiments · Wilson score intervals · Power analysis · Forward-in-time validation · Bias detection · Backtesting
Testing	Catch2 · pytest · Property/randomized testing · Load testing · Containerized CI suites

Featured Work

Eight self-directed builds. Every metric below is reproduced from a test suite, a benchmark harness, or a reported confidence interval.

DistributedCache

C++17 · POSIX Sockets · CMake · Catch2

A 4,700-line in-memory key-value cache with zero runtime dependencies.

O(1) LRU eviction — hash map plus a hand-written doubly linked list
Min-heap TTL expiry, verified across 62 tests / 1.19M assertions
Traced a 5.25× read-throughput collapse to cache-line contention on a shared mutex; 16-way lock striping returned 5.37× at 8 threads (615K → 3.30M ops/sec)
Consistent hashing over 200 virtual nodes — key remapping capped at 24.4% on cluster growth vs. 75% for modulo hashing

NameForge

Python · PostgreSQL · FastAPI · UDP · Docker

A 6,878-line DNS resolution backend that hand-encodes the binary wire format — no DNS library.

Iterative and recursive resolution across a simulated root / TLD / authoritative hierarchy
Composite B-tree index cut database statements 99.3% (4,559 → 33)
157× faster lookups at 200K records (5.486 ms → 0.035 ms)
16-endpoint REST API sustained 1,981 req/sec across four workers with zero failures

CollabSync

Node.js · WebSockets · PostgreSQL · React · CI/CD

A real-time collaborative editor with an Operational Transform engine written from scratch.

Concurrent editors converge byte-identically — proven across 30,000 randomized multi-client scenarios
Materialized snapshot + append-only operation log under compare-and-set versioning
Reconnection cost cut from O(missed ops) to O(1) messages via idempotent replay-then-snapshot
0.035 ms median edit round-trip · 153-test containerized CI suite

Pramaan

Backend Systems · Trust & Safety · Agentic AI

Trust-and-safety infrastructure for AI agents — Sanskrit pramāṇa, "proof."

341 tests passing, reproducible from a clean clone with make all
Built for Razorpay AI Buildathon — Track 01: AI Growth & Agentic Commerce
Sole architect and engineer: design, implementation, testing, documentation

Vasool (वसूल) — Revenue Recovery Agent

Python · scikit-learn · SQL · FastAPI · GitHub Actions

An AI revenue-recovery agent built for Razorpay AI Hackathon — Track 03.

Randomized 15% holdout over 2,000 payment failures worth ₹2.11 Cr
+14.2pp incremental recovery lift (95% CI +9.6 → +18.9), Wilson intervals, pre-registered power analysis
Caught two silent measurement biases that had hidden a 2.3× overstatement — reported lift moved −3.0pp → +14.2pp
Three-tier root-cause classifier over a 19-class taxonomy at 91.0% precision, ₹0.10 cost per ₹1 recovered
510-test CI pipeline reproduces every published figure

NEXUS AI

PyTorch · FinBERT · LangChain · FAISS · Docker

A hybrid financial forecasting and decision-support agent.

LSTM over 1,460-day price histories + technical indicators + FinBERT news sentiment — R² = 0.92, RMSE ≈ 2.5%
60-day rolling backtest reports every prediction as an uncertainty band that widens under volatile regimes
RAG layer (LangChain + FAISS + Gemini) turns 200-page filings into source-cited recommendations
Dockerized Streamlit dashboard

Semantic Search Engine

Sentence-Transformers · FAISS · GMM · FastAPI

Production-shaped vector retrieval over 19,125 documents with a custom semantic cache.

BIC-optimized GMM (k=30) with PCA 384D → 50D, hitting 88.8% high-confidence cluster assignments
Two-layer semantic caching with cluster-aware routing — 1,297× latency speedup (28 ms → 22 μs)
Containerized, runs locally and in Docker

Loan Performance Intelligence Engine

LightGBM · scikit-learn · Pandas · SQL

Portfolio risk forecasting across 268,125 monthly records over 9,885 loans.

Macro F1 +30% (0.408 → 0.530) under strictly forward-in-time validation
Five prediction heads calibrated with isotonic and Platt scaling — Brier score cut 46–82%, so probabilities drive loss provisioning, not just ranking
Traced a 0.23% "clean records" score to structurally null columns miscounted as missing; batch integrity recovered to 92.14/100

_{Also in the workshop: a local AI chat assistant serving Gemma-4-12B on Apple Silicon through an OpenAI-compatible API at 15–17 tok/sec, plus CNN and regression notebooks from the ML fundamentals track.}

Experience

LLM Engineering Intern · amasQIS.ai (Remote)

Mar 2025 – Sept 2025

Architected a four-stage decoupled pipeline — preprocessing → fine-tuning → evaluation → inference — with file-based handoff and zero shared state, so any stage reruns independently after a failure instead of restarting the workflow.
Engineered preprocessing over 50,000+ customer records with null auditing, stratified splitting and identity-leakage checks — trimmed data noise 40% and lifted recommendation relevance 15% over a rules-based baseline.
Fine-tuned an 8B reasoning model (DeepSeek-R1-Distill-Llama-8B) with 4-bit quantized LoRA on a single GPU.
Hardened multi-epoch GPU training with 100-step checkpointing and fixed seeds, bounding interruption cost to one checkpoint interval, and shipped a typed JSON contract consumed directly by downstream services.
Established the evaluation harness that decided which checkpoint reached the serving pipeline.

By the Numbers


354	9.01	1.19M	8
LeetCode problems solved	CGPA / 10	test assertions in one repo	shipped systems
3.30M	1,297×	₹2.11 Cr	Top 5
cache ops/sec achieved	cache speedup	payment failures analyzed	national hackathon finish

Signal

_{354 solved — 180 Medium, 152 Easy, 22 Hard. Medium outnumbers Easy, which is the
ratio I've been deliberately pushing toward.}

_{Contribution history is public and speaks for itself — the engineering detail lives in the repos above.}

Credentials

Certification	Issuer
Microsoft Certified: Fabric Data Engineer Associate	Microsoft
Microsoft Certified: Azure Data Fundamentals	Microsoft
Oracle Cloud Infrastructure 2025 — AI Foundations Associate	Oracle
Oracle Cloud Infrastructure 2025 Certified Foundations Associate	Oracle
Google IT Support Professional Certificate	Google
Applied Machine Learning in Python	University of Michigan · Coursera
Generative AI using IBM watsonx	IBM

Beyond the Code

Top 5 finish at HackSethu — a national-level hackathon — delivering a working solution under real-world constraints.
Core Member, Google Developer Group (GDG) — organized a college-fest hackathon end to end and drive community technical initiatives.
National Cadet Corps cadet — structured drills and camps; where the discipline behind the benchmarking habit actually came from.
354 LeetCode problems solved — 180 Medium, 152 Easy, 22 Hard — across arrays, trees, graphs, hashing, recursion and dynamic programming.

Education

Vellore Institute of Technology, Bhopal — B.Tech, Computer Science (AI & ML) · CGPA 9.01/10 · Sept 2023 – May 2027

Birla School, Pilani — CBSE AISSCE (Class XII) · 93.6% · 2022

Let's Build Something

I'm open to SDE and Machine Learning internships — especially teams that care about the measurement as much as the model.

"Benchmark first. Claim second."

Navigation Menu